Time Series Analysis: Forecasting and Control
Time Series Analysis: Forecasting and Control
Approximation of nonlinear systems with radial basis function neural networks
IEEE Transactions on Neural Networks
Hi-index | 0.00 |
The classical autoregressive moving average model (ARMA) fails to satisfy the high request for precision in predicting nonlinear and nonstationary systems. Overcoming the difficulty, a hybrid prediction method is proposed in this paper, which organically couples the radial basis function prediction neural network (RBFPNN) and the functional-coefficient autoregressive prediction model (FARPM). An observation time series characterized by nonlinearity and nonstationarity can be technically decomposed with the wavelet analysis tool into two clusters of sequences, i.e. the smooth sequences and the stationary sequences, which can be effectively predicted with RBFPNN and FARPM respectively. Then, the integrated prediction is obtained by merging the results of RBFPNN and FARPM. It's indicated by the simulation that the prediction precision for one step, 4 steps and 12 steps can be improved at least by 41%, 60% and 60% respectively, compared to the prediction with ARMA, RBFPNN and FARPM separately.